Data Characteristics
Pharmacovigilance data for psychiatric disorders originates from clinical trial reports, real-world evidence (RWE), physician case reports, patient self-reporting systems, and academic literature. Data updates frequently, especially after new drug launches or new adverse event reports. Document structures typically follow standardized reporting templates, such as CIOMS I forms or MedWatch 3500A forms. These forms are highly structured and include fields like patient demographics, medication history, adverse reaction descriptions, onset time, severity, and outcome. Adverse reaction descriptions often involve subjective feelings, behavioral changes, and mental status assessments. This information is usually in free-text format and may contain medical terminology, slang, or non-standardized patient expressions. Dosage units are typically milligrams (mg) or milliliters (mL), and frequency units are per day or per week.
Constraints on Multi-turn Conversations and Prompts
Diverse data sources and frequent updates require the multi-turn conversation system to quickly absorb new information and update its knowledge base. This ensures prompt references use timely information. The coexistence of structured reports and free text means prompt design must precisely extract structured fields and semantically understand unstructured text. For example, for free-text descriptions of mental status, the system needs to clarify ambiguous information through multi-turn questioning to improve the accuracy of adverse reaction judgments. Due to the nature of psychiatric disorders, patient or reporter descriptions of adverse reactions can be highly personal and subjective. Prompts must remain neutral during conversations, avoid leading questions, and identify and process emotional tendencies to obtain objective facts. The presence of medical terminology, slang, and non-standardized expressions challenges model comprehension and prompt robustness. This may require pre-defined synonym libraries or explanatory guidance within prompts.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000 tokens | Psychiatric adverse event reports often include detailed medical history and multi-turn symptom descriptions, requiring a long context window for conversational coherence. |
Chunk size (Chunk Size) | 500 characters | Ensures knowledge base chunks contain enough information to understand complex symptom descriptions, while avoiding overly long chunks that impact recall efficiency. |
Recall count (Recall Count) | Top 8 entries | Given the complexity and diversity of adverse reactions in psychiatric disorders, increasing the recall count covers more potentially relevant information. |
Similarity threshold (Similarity Threshold) | 0.75 | For potential semantic ambiguity in psychiatric disorder descriptions, a higher threshold ensures strong relevance of recall results and reduces noise. |
Rerank result count (Reranked Return Count) | Top 5 entries | Further filters initial recall results to focus on the most relevant items, improving accuracy for the end-user. |
systemPrompt | Calibrate by testing | Needs to include clear role-setting (e.g., "I am a pharmacovigilance assistant"), task objectives ("collect and analyze adverse reaction information"), and emphasize neutrality and objectivity. |
Common Pitfalls
- Model prompts
connection errorduring a conversation: This usually results from restricted model API access or network instability. Check network proxy configurations and API key validity. - Knowledge base content includes images, but the model cannot display them in the conversation: The prompt does not explicitly instruct the model on how to handle image resources. The model cannot directly render images by default. Guide the model via prompts to return image links or instruct the user to view the original document.
- Inaccurate extraction of key variables like medication dosage and frequency from the conversation: Prompt design is too broad and does not specify the exact fields and data formats to extract, leading to inaccurate understanding or extraction failure by the model.
Verification Steps
- Test if multi-turn conversations accurately identify and extract drug names, dosages, administration routes, and adverse reaction onset times from reports.
- Verify if the system can guide users to provide more specific symptom information through follow-up questions when encountering vague or non-standardized descriptions (e.g., "feeling unwell," "low mood").
- Examine knowledge base recall results. Confirm that entering specific adverse reaction symptoms accurately matches relevant drug instructions or previous cases. Verify that recall count and similarity meet expectations.
- Simulate a user submitting a complex report containing various psychiatric disorder symptoms. Evaluate if the system can collect necessary information completely in a single conversation or effectively guide information supplementation in multi-turn conversations.
Note: The values provided are common starting points. Measure them against your own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.